Small and micro enterprise credit prediction method and system based on reinforcement learning

Abstract

The invention provides a reinforcement learning-based small and micro enterprise credit prediction method and system, and the method comprises the steps: obtaining the credit evaluation data of a loan application customer; evaluating a credit score of the loan application customer by using a credit scoring model; giving a classification result according to a relationship between the credit score and a score threshold value; using the trained reinforcement learning model to optimize a score threshold according to the environment state; the environment state comprises actual repayment conditions of customers after credit expiration, credit scores and loan information of repayment customers and non-repayment customers, and can also comprise other state characteristics which can be used for describing market conditions, such as the overall bad debt rate of the industry. According to the scheme, a reinforcement learning method is introduced for the scoring model, the scoring threshold value of credit scoring can be dynamically optimized, and compared with a traditional method, the real-time updating mechanism based on actual repayment data can enable the scoring threshold value to be set more reasonably and timely, so that the scoring threshold value is not too high or too low at any time, and the optimal input-output ratio is achieved.

Description

Technical Field

[0001] The present invention belongs to the field of computer technology, and in particular to a method and system for predicting credit of small and micro enterprises based on reinforcement learning. Background Art

[0002] Small and micro enterprises are often constrained by funding constraints in the process of growth and technological innovation. In the context of limited capital supply in the credit market, the financing difficulties faced by small and micro enterprises are particularly serious. For a long time, financing difficulties have been one of the main obstacles to the development of small and medium-sized enterprises, among which inappropriate credit risk assessment mechanisms play a key role.

[0003] In order to solve the financing problem, it is very important to build a reasonable and effective credit evaluation system for small and micro enterprises. In this process, two evaluation methods have been developed: expert evaluation method and model evaluation method.

[0004] Initial credit assessment relies on expert evaluation, a widely used methodology. This method uses a series of rules and indicators to evaluate credit status based on historical data and expert experience. Expert evaluation can deeply analyze the financial and non-financial information of enterprises, but it has the following disadvantages: it relies on personal experience and lacks scientific and normative nature; it is difficult to standardize and widely promote; it lacks flexibility and is difficult to adapt to environmental changes.

[0005] With the continuous segmentation and specialization of the loan market, various commercial banks have launched loan products for small and micro enterprises, which has led to an increasing professional requirement for pre-loan review of small and micro enterprises, and at the same time, higher standards have been set for the scientificity and practicality of credit evaluation methods. In the academic community, researchers have conducted extensive and in-depth discussions on the credit model, rating indicators, rating methods and default probability prediction of small and micro enterprise loan business based on the needs of practical operations, and have accumulated rich research results, in order to provide theoretical support and practical guidance for solving the financing difficulties of small and micro enterprises.

[0006] Based on the above background, in the credit approval practice of commercial banks, the use of special credit scoring models to comprehensively evaluate the information provided by enterprises applying for loans in order to predict the future default risk of enterprises has become a standard and crucial pre-loan review step. Given that this process has a fundamental impact on bank credit decisions and risk management, the effectiveness and accuracy of the scoring model naturally become the focus of common concern and optimization for all commercial banks. By continuously optimizing the scoring model, commercial banks can improve the efficiency of loan approval for small and micro enterprises, ensure the efficient allocation of credit resources and minimize risks. To this end, various banks have proposed methods for implementing scoring models, such as a risk identification method, device and system disclosed in a Chinese patent [Publication No.: CN110363449B]. However, among the various scoring models currently available, there is an outstanding problem that has not been solved. The scoring model needs a scoring threshold to make a prediction. At present, the determination of this scoring threshold still relies on expert experience and cannot be updated in a timely manner. There is a problem of lagging evaluation model updates. Summary of the invention

[0007] The purpose of the present invention is to provide a small and micro enterprise credit prediction method based on reinforcement learning to address the above problems;

[0008] Another object of the present invention is to provide a small and micro enterprise credit prediction system based on reinforcement learning for executing the aforementioned method.

[0009] In order to achieve the above object, the present invention adopts the following technical solutions:

[0010] A small and micro enterprise credit prediction method based on reinforcement learning, the method comprising:

[0011] Obtain credit rating data of loan applicants;

[0012] Use the trained credit scoring model to assess the credit scores of loan applicants;

[0013] Give classification results based on the relationship between credit score and score threshold;

[0014] Utilizing the trained reinforcement learning model, optimizing the scoring threshold according to the state of the environment;

[0015] The environmental status includes the actual repayment status of customers after the credit expires, as well as the credit scores and loan information of customers who have repaid and those who have not repaid. It can also include other status characteristics that can be used to describe market conditions, such as the overall bad debt rate of the industry.

[0016] In the above-mentioned reinforcement learning-based small and micro enterprise credit prediction method, the credit evaluation data includes the personal data of the business owner of the loan applicant, the business operation data, and the industry environment data;

[0017] Scoring models assess the credit scores of loan applicants in the following ways:

[0018] Receive credit evaluation data of loan applicants;

[0019] The credit evaluation data is divided into business owner personal data, business operation data, and industry environment data;

[0020] Input three types of data, namely, business owner’s personal data, business operation data, and industry environment data, into the scoring model;

[0021] The scoring model evaluates credit scores based on three types of data.

[0022] In the above-mentioned reinforcement learning-based small and micro enterprise credit prediction method, when the credit score is less than or equal to the score threshold, the current loan applicant is classified as a risky customer, with a high probability of not repaying the loan, and it is not recommended to lend. Otherwise, the current loan applicant is classified as a high-quality customer with a high probability of repaying the loan, and it is recommended to lend. Bank staff decide whether to lend based on the classification results of the current loan applicant.

[0023] In the above-mentioned reinforcement learning-based credit prediction method for small and micro enterprises, the reinforcement learning model is trained to give a scoring threshold for the next state under a given state so as to optimize the total profit of the loan application.

[0024] In the above-mentioned reinforcement learning-based small and micro enterprise credit prediction method, the action of the reinforcement learning model is the scoring threshold determined by the agent in a given state; it can also be considered as the difference from the scoring threshold of the previous state.

[0025] The rewards of the reinforcement learning model include the reduction of profits and / or losses due to the increase of the scoring threshold, or the increase of profits and / or losses due to the reduction of the scoring threshold. That is, when the scoring threshold increases, the lending threshold rises, the corresponding bad debts will also decrease, and the losses will therefore decrease, and the total profit will increase; but at the same time, the profit brought by the part of customers who may repay the loan will also decrease due to the reduction of the loan amount, so the profit will decrease, and the total profit will also decrease accordingly; therefore, the result of the increase of the scoring threshold may be an increase in the total profit, or a decrease in the total profit. If the profit decreases more, the total profit will decrease, and if the loss decreases more, the total profit will increase.

[0026] Similarly, as the scoring threshold is lowered, the lending threshold is lowered, the corresponding bad debts will also increase, the losses will increase, and the total profit will decrease; at the same time, the increase in the amount of lending will increase the profits brought by these customers with lower scores, so the total profit will increase accordingly; therefore, the result of lowering the scoring threshold may be an increase or a decrease in the total profit. If the profit increases more, the total profit will increase; if the loss increases more, the total profit will decrease.

[0027] The state of the reinforcement learning model includes the current state scoring threshold, the loan customer's loan information, credit score and repayment status.

[0028] In the above-mentioned reinforcement learning-based credit prediction method for small and micro enterprises, historical loan data is used to train the reinforcement learning model;

[0029] Historical loan data includes several loan cases including the credit score, loan principal, loan interest rate, repayment method, and repayment status of the loan customer. The scoring model is trained first, and then the reinforcement learning is trained. The credit score is predicted using the trained scoring model.

[0030] In the above-mentioned reinforcement learning-based small and micro enterprise credit prediction method, the reinforcement learning model is trained in the following way:

[0031] Construct a data tuple for each loan case according to the scoring score - profit / loss;

[0032] Initialize the value of the action value function of reinforcement learning to 0 and determine an initial scoring threshold;

[0033] Calculate the total profit under the initial scoring threshold;

[0034] Keep trying new scoring thresholds until you get the best scoring threshold, under which the total profit of all loan cases is optimal;

[0035] In this process, the action-value function of the reinforcement learning model is continuously updated to better estimate the expected reward of taking a specific action in a specific state.

[0036] In the above-mentioned reinforcement learning-based small and micro enterprise credit prediction method, during the training process, if the current state scoring threshold t AT Scoring threshold relative to the previous state (t-1) AT Decreases, then the score is calculated based on the loan customer's repayment situation at (t+1) AT -t AT The total profit of the loan customers between them. If the total profit is greater than the first set value, the reward is negative, otherwise the reward is positive;

[0037] If the current state scoring threshold t AT Relative to the previous state (t-1) AT Increase, then the score is calculated based on the loan customer's repayment situation at t AT -(t+1) AT If the total profit is greater than the second set value, the reward is positive, otherwise the reward is negative.

[0038] In the above-mentioned reinforcement learning-based small and micro enterprise credit prediction method, during the reasoning process, the scoring threshold optimization work is performed periodically according to the repayment period of different loan customers, and in each cycle, the scoring threshold is optimized according to the actual repayment situation, credit score, and loan information of the current repaying customer.

[0039] A small and micro enterprise credit prediction system based on reinforcement learning, used for performing credit prediction through the above method, including a data acquisition module, a credit scoring module, a credit decision module, and a threshold update module;

[0040] Data acquisition module, used to obtain credit evaluation data of loan applicants;

[0041] A credit scoring module, used to evaluate the credit score of the loan applicant based on the credit evaluation data;

[0042] The credit decision module is used to make a classification decision on whether to lend based on the relationship between the credit score and the score threshold;

[0043] A threshold updating module, used for optimizing the scoring threshold according to the environmental status;

[0044] The environmental status includes the customer's actual repayment status after the credit expires, as well as the credit scores and loan information of the repaying customers and the non-repaying customers.

[0045] The advantages of the present invention are:

[0046] 1) This solution introduces a reinforcement learning method into the scoring model, which can dynamically optimize the scoring threshold of the credit score. Compared with the traditional method, this real-time update mechanism based on actual repayment data can make the scoring threshold setting more reasonable and timely, so that it will not be too high or too low at any time, achieving the best input-output ratio;

[0047] 2) The scoring model proposed in this scheme describes the nonlinear relationship between various factors and default, including individual, enterprise and industry data, providing a comprehensive credit assessment perspective, which helps to more accurately capture the credit status of small and micro enterprises;

[0048] 3) This solution uses historical loan cases to construct score-profit / loss data tuples to train reinforcement learning models, with the expected profit maximization as the optimization goal. By comprehensively considering the profits of good customers and the losses of bad customers, it balances risks and returns when approving loans, avoids overly conservative or overly risky decisions, and helps banks increase profits under the premise of controllable risks;

[0049] 4) Based on the characteristics of small and micro enterprises, this solution classifies the data and conducts multi-perspective evaluation. At the same time, the adjustment of the scoring threshold is related to the customer's actual repayment situation and expected profit, and the optimization goals of the model are consistent with the business goals, avoiding the separation of the effects of the model and engineering implementation. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 A flowchart of the training and reasoning process of the scoring model in the small and micro enterprise credit prediction method based on reinforcement learning of the present invention;

[0051] Figure 2 A working principle diagram of reinforcement learning in the small and micro enterprise credit prediction method based on reinforcement learning of the present invention;

[0052] Figure 3 It is a flow chart of a method for constructing a reinforcement learning model in a small and micro enterprise credit prediction method based on reinforcement learning of the present invention;

[0053] Figure 4 It is a flow chart of the reasoning process of the small and micro enterprise credit prediction method based on reinforcement learning of the present invention;

[0054] Figure 5 This is a diagram illustrating the loan approval situation of a certain bank within three months in the embodiment. DETAILED DESCRIPTION

[0055] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0056] This solution discloses a reinforcement learning-based credit prediction method for small and micro enterprises and a system for implementing the method.

[0057] like Figure 1 As shown in the figure, firstly, a scoring model is constructed. This scoring model adopts a neural network model or a support vector machine model. Using the labeled credit evaluation data set, it is divided into a training set and a test set in a ratio of about 9:1 to train a credit evaluation model for small and micro enterprises. The credit evaluation data includes the personal data of the business owner of the loan applicant, the business operation data, and the industry environment data.

[0058] As shown, the business owner's personal data may include any one or more combinations of identity information, educational background, professional experience, credit record, asset status, income status, etc.

[0059] Business operating data may include any one or more combinations of financial statements, operating income, profitability, assets and liabilities, liquidity indicators, inventory and accounts receivable, and years of operation.

[0060] Industry environment data may include industry trends (industry growth trends, market size, etc.), competition conditions (degree of competition within the industry, market share distribution, etc.), policy environment (government support policies for the industry, regulatory restrictions, etc.), economic cycles (macroeconomic cycles, industry cycles, etc.), supply chain conditions (supplier stability, raw material costs, etc.), market demand (changes in market demand, consumer preferences, etc.), and any one or more combinations thereof.

[0061] When the credit evaluation data is obtained, it is cleaned and some data without actual value is filtered out. Then, the data is classified into business owner personal data, business operation data, and industry environment data. In other words, a series of credit evaluation data will be collected for each loan applicant, and the data will eventually be divided into three groups. Subsequently, the credit evaluation data of the loan applicants are grouped and sent to the model. There will be score labels in the training stage, and the model will output the score results based on the credit evaluation data in the inference stage.

[0062] In general, the overall idea of ​​credit scoring is to estimate the probability y of a bad loan based on the feature vector of the loan applicant x, that is, various credit evaluation data:

[0063]

[0064] Estimated probability P r This is then converted into the applicant's credit score. CS Finally, we select the credit score threshold t AT to make a decision; if s CS ≥t AT If the score is chosen properly, then it can be distinguished between good customers and bad customers.

[0065] The final stage of building a credit scoring model is to optimize the scoring threshold t AT. Generally speaking, the purpose of selecting the best cut-off point is to minimize the classification error rate of loan applications. However, when determining the threshold, in addition to the basic classification method, a variety of external factors need to be considered comprehensively, including but not limited to the actual misclassification cost, the decision maker's utility evaluation function, possible biases in the sample selection process, and the dynamic changes of the data population. To this end, this scheme proposes to introduce a reinforcement learning model to optimize the scoring threshold in credit assessment. Reinforcement learning is trained to give the scoring threshold of the next state under a given state so that the sum of profits of all loan applications in a specific time period is optimal.

[0066] The specific method is as follows:

[0067] like Figure 2 As shown, in this scheme, the action of the reinforcement learning model adopted is the scoring threshold determined by the agent in a given state, which can also be considered as the difference from the scoring threshold of the previous stage. The rewards of the reinforcement learning model include the reduction of profits and / or losses due to the increase of the scoring threshold, or the increase of profits and / or losses due to the reduction of the scoring threshold, where the increase or decrease of the scoring threshold is the increase or decrease of the scoring threshold relative to the previous state. When the changes in profits and losses show a decrease in profits or an increase in losses on the whole, the total profit is reduced and the environment gives negative feedback; when the profits increase or the losses decrease on the whole, the total profit is increased and the environment gives positive feedback.

[0068] The state of the reinforcement learning model includes the current state scoring threshold, loan information of each loan customer, such as loan principal, interest rate, loan period, repayment method, and repayment status, etc. These states can describe the current characteristics of the credit portfolio, market conditions, and economy, so that the intelligent agent can make scoring threshold determination actions that are in line with the current market and realistic.

[0069] This system uses the Bellman equation to calculate the action value function, which is defined as follows:

[0070] Q(S t , A t )←Q(S t , A t )+α[R t +γmax a Q(S t+1 , a)-Q(S t , A t )](2)

[0071] Where γ is the discount rate, a value between 0 and 1 that is used to calculate the present value of future rewards, reflecting the decline in the value of future rewards relative to immediate rewards;

[0072] α is the learning rate;

[0073] R t +γmax a Q(S t+1 , a)-Q(S t , A t ) is the approximate error.

[0074] Q(S t , A t ) is the action value function, predicting from state S t Start with action A t The expected return that can be obtained by following the optimal strategy is the action value function in the reinforcement learning method. It reflects the expected value of the cumulative reward that the agent can obtain by taking a certain action in a certain state. It is the concept of cumulative value, and the initial value is generally 0. By initializing the Q value to 0, an unbiased and neutral starting point is constructed for the learning algorithm, ensuring the fairness and universality of the learning process.

[0075] R t : The immediate reward after taking action At at time t, the feedback received from the environment, this solution calculates the real profit situation based on the current score threshold and the actual repayment situation of small and micro enterprises with credit scores greater than the score threshold, through the loan principal, loan interest rate, loan time, repayment method and time, etc. For example, if 5 loans are issued at the current acceptance threshold, after maturity, 4 loans repay the principal and interest on time, and one of them does not repay on time, the real profit = the total profit of the 4 loans - the loss of the unpaid loan.

[0076] If the current state scoring threshold t AT Scoring threshold relative to the previous state (t-1) AT Decreases, then the score is calculated based on the loan customer's repayment situation at (t+1) AT -t AT The total profit of loan customers between (since it is a training phase and historical data is used, this batch of loan cases may exist). If the total profit is greater than the first set value, the reward is negative, otherwise the reward is positive.

[0077] If the current state scoring threshold t AT Scoring threshold relative to the previous state (t-1) AT Increase, then the score is calculated based on the loan customer's repayment situation at t AT -(t+1) AT If the total profit is greater than the second set value, the reward is positive, otherwise the reward is negative.

[0078] The first set value and the second set value may be the same or different. For example, the first set value may be 0, and the second set value may be 10,000.

[0079] max a Q(S t+1 ,a),For the next state St+1, the maximum value of the action value function Q of all possible actions a represents the agent’s maximum expectation of future rewards.

[0080] The design of the above actions, states, and rewards realizes the construction of a cost-sensitive optimization strategy by combining the profit maximization principle and the instance-based misclassification cost framework, thereby improving the economic benefits and efficiency of loan approval.

[0081] like Figure 3 As shown, during the training process, historical loan data is used to train the reinforcement learning model. The historical loan data includes several loan cases containing information such as the credit score of the loan customer (predicted using the trained scoring model), loan principal, loan interest rate, loan time, repayment method, and whether to repay the loan upon maturity.

[0082] Construct a data tuple for each loan case by credit score score-profit / loss;

[0083] Profit = interest * cash cost ratio * loan principal; loss = principal * (1 + cash cost ratio)

[0084] Here is a simple repayment method. In actual situations, there are more complicated repayment methods, such as installment repayment after the loan, or the lender only repays a few installments. Other loan methods calculate profits or losses according to specific circumstances, which will not be elaborated here.

[0085] The value of the action value function of reinforcement learning is initialized to 0, and an initial scoring threshold is determined. A relatively small scoring threshold can be determined empirically by those skilled in the art.

[0086] The historical loan data (including the individual loan case data of the data tuple) is divided into batches to train the reinforcement learning model. Each batch is trained as follows:

[0087] Calculate the total profit under the initial scoring threshold for the first batch of training data;

[0088] Keep trying new scoring thresholds until the best scoring threshold is obtained; under this best scoring threshold, the total profit of this batch of loan cases is optimal.

[0089] In this process, the action-value function and strategy of the reinforcement learning model are continuously updated to better estimate the expected return of taking a specific action in a specific state and to be able to make actions that obtain the optimal profit and score threshold.

[0090] Then continue to use the second and third batches of training data to repeat the above training process. When training subsequent batches, the scoring threshold can use the initial scoring threshold or the best scoring threshold finally determined in the previous batch.

[0091] For example, there are 200 loan cases in total, which are divided into 5 batches, each with 40 loan cases. First, the reinforcement learning model is trained using the data of the first batch of loan cases to update the action value function and strategy; then the reinforcement learning model is trained using the second, third, fourth, and fifth batches, and finally it is able to make the optimal scoring threshold decision based on the latest status.

[0092] This system adopts the radial basis function (RBF) network as an approximation tool for the value function. Its advantage is that it can create a more flexible and changeable function form, which perfectly meets the requirement of the value function to present a bell-shaped distribution in the state and action space. Compared with other complex neural network models, the RBF network shows a faster convergence speed, and is easier to set up and maintain, and the computational overhead is significantly reduced. Based on this, the state variables are processed through the RBF transformation to generate a total of 2,000 feature quantities, which are then normalized and applied as input data to the stochastic gradient descent model for further analysis and optimization.

[0093] Finally, if Figure 4 As shown in the figure, the trained scoring model and reinforcement learning model of this solution are deployed in the credit business environment, and the reasoning process is as follows:

[0094] Obtain credit rating data of loan applicants;

[0095] Use the trained credit scoring model to assess the credit scores of loan applicants;

[0096] Give classification results based on the relationship between credit score and score threshold;

[0097] Using the trained reinforcement learning model, the scoring threshold is optimized according to the environmental status. Specifically, the scoring threshold is optimized according to the customer's actual repayment situation after the credit expires, the credit score of the repaying customer, the credit score of the non-repaying customer, and the loan information of each loan customer.

[0098] During the deployment process, a scoring threshold can be determined based on experience or historical data. Then, this system is used to perform credit scoring based on this scoring threshold. After a period of time, a group of customers will be granted credit loans because their credit scores are higher than the scoring threshold. These customers have different loan terms and repayment methods. Each customer can be taken as a tuple, and the scoring threshold can be updated once according to the customer's repayment situation. It is also possible to take a period of time covering the repayment time of several loan customers as one period, and update the scoring threshold once according to the repayment situation of the customers within this period.

[0099] Taking the latter as an example, assuming that the current scoring threshold T0=650, with one month as a period, the scoring threshold is optimized once at the end of each month.

[0100] like Figure 5 As shown, suppose the bank approved 100 loans to people with credit scores above 650 in one month, 80 in the second month, and 70 in the third month.

[0101] Suppose that among 100 loans, 2 have a repayment period of 1 month, 2 have a repayment period of 3 months, 2 have a repayment period of 6 months, and the rest have a repayment period of more than 1 year. Among 80 loans, 3 have a repayment period of 1 month, 2 have a repayment period of 3 months, and the rest have a repayment period of more than 1 year. Among 70 loans, 5 have a repayment period of 3 months, 4 have a repayment period of 6 months, and the rest have a repayment period of more than 1 year.

[0102] The bank records the detailed information of each loan, including the principal, interest rate, repayment period, etc. There should be two repayments in the first loan cycle (i.e. the first month). Then, one month later, the reinforcement learning model will update the scoring threshold based on the repayment status of the two loans and the credit scores, loan terms, loan amounts, loan interest and other environmental conditions of the two loan customers. The scoring threshold may be maintained, higher or lower. For example, if both loans are repaid on time and the credit scores of both loans are close to 650, the scoring threshold for the next cycle may be optimized to a value less than 650. If one loan is not repaid on time, the scoring threshold for the next cycle may be optimized to a value greater than 650. Whether to update and how much to update is determined based on the acquired environmental conditions, as well as the strategies and action value functions learned by reinforcement learning during the training phase.

[0103] There should be 3 repayments in the second loan cycle (i.e. the second month), 2 in the third loan cycle (i.e. the third month), 2 in the fourth loan cycle, 7 in the fifth loan cycle... The above method is repeated for each loan cycle to update the scoring threshold.

[0104] Through the above process, this solution, on the one hand, can make credit scores for loan applicants based on historical financial, personal and industry data, and can comprehensively and quantitatively evaluate the credit quality of small and micro enterprises, including non-financial factors, thus improving the comprehensiveness and accuracy of credit assessment. On the other hand, the scoring threshold is updated in a timely manner in stages according to the repayment period of different customers, which is in line with the characteristics of different loan periods in various loan cases in the credit scenario. At the same time, the acceptance threshold can be dynamically adjusted according to the feedback of actual user overdue situations, avoiding the problem of setting too high or too low, and achieving the best input-output ratio.

[0105] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

[0106] Although the terms such as data acquisition module, credit scoring module, credit decision module, threshold update module, etc. are frequently used in this article, the possibility of using other terms is not excluded. These terms are used only to more conveniently describe and explain the essence of the present invention; interpreting them as any additional restrictions is contrary to the spirit of the present invention.

Claims

1. A small and micro enterprise credit prediction method based on reinforcement learning, characterized in that: The method includes: Obtain credit rating data of loan applicants; Use the trained credit scoring model to assess the credit scores of loan applicants; Give classification results based on the relationship between credit score and score threshold; Utilizing a trained reinforcement learning algorithm, optimizing the scoring threshold according to the state of the environment; The environmental status includes the customer's actual repayment status after the credit expires, as well as the credit scores and loan information of the repaying customers and the non-repaying customers.

2. The method for predicting small and micro enterprise credit based on reinforcement learning according to claim 1 is characterized in that: The credit evaluation data includes personal data of the business owner of the loan applicant, business operation data, and industry environment data; Scoring models assess the credit scores of loan applicants in the following ways: Receive credit evaluation data of loan applicants; The credit evaluation data is divided into business owner personal data, business operation data, and industry environment data; Input three types of data, namely, business owner’s personal data, business operation data, and industry environment data, into the scoring model; The scoring model evaluates credit scores based on three types of data.

3. The method for predicting small and micro enterprise credit based on reinforcement learning according to claim 1 is characterized in that: When the credit score is less than or equal to the score threshold, the current loan application customer is classified as a risky customer, otherwise, the current loan application customer is classified as a high-quality customer.

4. The method for predicting small and micro enterprise credit based on reinforcement learning according to claim 1 is characterized in that: The reinforcement learning algorithm is trained to give a score threshold for the next state in a given state so as to optimize the total profit of the loan application.

5. The method for predicting small and micro enterprise credit based on reinforcement learning according to claim 4 is characterized in that: The actions of the reinforcement learning algorithm are, in a given state, the scoring threshold determined by the agent; The rewards of the reinforcement learning algorithm include a reduction in profit and / or loss due to an increase in the scoring threshold, or an increase in profit and / or loss due to a decrease in the scoring threshold; The state of the reinforcement learning algorithm includes the scoring threshold of the current state, the loan information, credit score and repayment status of the loan customer.

6. The method for predicting small and micro enterprise credit based on reinforcement learning according to claim 5 is characterized in that: Use historical loan data to train a reinforcement learning algorithm; The historical loan data includes several loan cases including the loan customer’s credit score, loan principal, loan interest rate, repayment method, and repayment status.

7. The method for predicting small and micro enterprise credit based on reinforcement learning according to claim 6 is characterized in that: The reinforcement learning algorithm is trained as follows: Construct a data tuple for each loan case according to the scoring score - profit / loss; Initialize the value of the action value function of reinforcement learning to 0 and determine an initial scoring threshold; Calculate the total profit under the initial scoring threshold; Keep trying new scoring thresholds until the best scoring threshold is obtained; During this process, the action-value function of the reinforcement learning algorithm is continuously updated to better estimate the expected reward of taking a specific action in a specific state.

8. The method for predicting small and micro enterprise credit based on reinforcement learning according to claim 7 is characterized in that: During the training process, if the current state scoring threshold t AT Scoring threshold relative to the previous state (t-1) AT Decreases, then the score is calculated based on the loan customer's repayment situation at (t+1) AT -t AT The total profit of the loan customers between them. If the total profit is greater than the first set value, the reward is negative, otherwise the reward is positive; If the current state scoring threshold t AT Relative to the previous state (t-1) AT Increase, then the score is calculated based on the loan customer's repayment situation at t AT -(t+1) AT If the total profit is greater than the second set value, the reward is positive, otherwise the reward is negative.

9. The method for predicting small and micro enterprise credit based on reinforcement learning according to claim 7 is characterized in that: During the reasoning process, the scoring threshold optimization work is performed periodically according to the repayment period of different loan customers, and in each cycle, the scoring threshold is optimized according to the actual repayment situation, credit score, and loan information of the current repayment customer.

10. A small and micro enterprise credit prediction system based on reinforcement learning, characterized in that: Used to perform credit prediction by the method described in any one of claims 1 to 9, comprising a data acquisition module, a credit scoring module, a credit decision module, and a threshold update module; Data acquisition module, used to obtain credit evaluation data of loan applicants; A credit scoring module, used to evaluate the credit score of the loan applicant based on the credit evaluation data; The credit decision module is used to make a classification decision on whether to lend based on the relationship between the credit score and the score threshold; A threshold updating module, used for optimizing the scoring threshold according to the environmental status; The environmental status includes the customer's actual repayment status after the credit expires, as well as the credit scores and loan information of the repaying customers and the non-repaying customers.

Images